File size: 58,580 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 |
{
"paper_id": "I08-1045",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:42:14.473060Z"
},
"title": "A Hybrid Feature Set based Maximum Entropy Hindi Named Entity Recognition",
"authors": [
{
"first": "Sujan",
"middle": [
"Kumar"
],
"last": "Saha",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Indian Institute of Technology Kharagpur",
"location": {
"postCode": "721302",
"region": "West Bengal India"
}
},
"email": "sujan.kr.saha@gmail.com"
},
{
"first": "Sudeshna",
"middle": [],
"last": "Sarkar",
"suffix": "",
"affiliation": {},
"email": "shudeshna@gmail.com"
},
{
"first": "Pabitra",
"middle": [],
"last": "Mitra",
"suffix": "",
"affiliation": {},
"email": "pabitra@gmail.com"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "We describe our effort in developing a Named Entity Recognition (NER) system for Hindi using Maximum Entropy (Max-Ent) approach. We developed a NER annotated corpora for the purpose. We have tried to identify the most relevant features for Hindi NER task to enable us to develop an efficient NER from the limited corpora developed. Apart from the orthographic and collocation features, we have experimented on the efficiency of using gazetteer lists as features. We also worked on semi-automatic induction of context patterns and experimented with using these as features of the MaxEnt method. We have evaluated the performance of the system against a blind test set having 4 classes-Person, Organization, Location and Date. Our system achieved a f-value of 81.52%.",
"pdf_parse": {
"paper_id": "I08-1045",
"_pdf_hash": "",
"abstract": [
{
"text": "We describe our effort in developing a Named Entity Recognition (NER) system for Hindi using Maximum Entropy (Max-Ent) approach. We developed a NER annotated corpora for the purpose. We have tried to identify the most relevant features for Hindi NER task to enable us to develop an efficient NER from the limited corpora developed. Apart from the orthographic and collocation features, we have experimented on the efficiency of using gazetteer lists as features. We also worked on semi-automatic induction of context patterns and experimented with using these as features of the MaxEnt method. We have evaluated the performance of the system against a blind test set having 4 classes-Person, Organization, Location and Date. Our system achieved a f-value of 81.52%.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Named Entity Recognition involves locating and classifying the names in text. NER is an important task, having applications in Information Extraction (IE), question answering, machine translation and in most other NLP applications.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "NER systems have been developed for English and few other languages with high accuracies. These systems take advantage of large amount of Named Entity (NE) annotated corpora and other NER resources. However when we started working on a NER system for Hindi, we did not have any NER annotated corpora for Hindi, neither did we have access to any comprehensive gazetteer list.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this work we have identified suitable features for the Hindi NER task. Orthography features, the suffix and prefix information, as well as information about the sorrounding words and their tags are used to develop a Maximum Entropy (MaxEnt) based Hindi NER system. Additionally, we have acquired gazetteer lists for Hindi and used these gazetteers in the Maximum Entropy (MaxEnt) based Hindi NER system. We also worked on semi-automatically learning of context pattern for identifying names. These context pattern rules have been integrated into the MaxEnt based NER system, leading to a high accuracy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The paper is organized as follows. A brief survey of different techniques used for the NER task in different languages and domains are presented in Section 2. The MaxEnt based NER system is described in Section 3. Various features used in NER are then discussed. Next we present the experimental results and related discussions. Finally Section 8 concludes the paper.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "A variety of techniques has been used for NER. The two major approaches to NER are: 1. Linguistic approaches.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Previous Work",
"sec_num": "2"
},
{
"text": "The linguistic approaches typically use rules manually written by linguists. There are several rulebased NER systems, containing mainly lexicalized grammar, gazetteer lists, and list of trigger words, which are capable of providing 88%-92% f-measure accuracy for English (Grishman, 1995; McDonald, 1996; Wakao et al., 1996) .",
"cite_spans": [
{
"start": 271,
"end": 287,
"text": "(Grishman, 1995;",
"ref_id": "BIBREF4"
},
{
"start": 288,
"end": 303,
"text": "McDonald, 1996;",
"ref_id": "BIBREF7"
},
{
"start": 304,
"end": 323,
"text": "Wakao et al., 1996)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Machine Learning based approaches.",
"sec_num": "2."
},
{
"text": "The main disadvantages of these rule-based techniques are that these require huge experience and grammatical knowledge of the particular language or domain and these systems are not transferable to other languages or domains.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Machine Learning based approaches.",
"sec_num": "2."
},
{
"text": "Machine Learning (ML) based techniques for NER make use of a large amount of NE annotated training data to acquire high level language knowledge. Several ML techniques have been successfully used for the NER task of which Hidden Markov Model (Bikel et al., 1997 ), Maximum Entropy (Borthwick, 1999 ), Conditional Random Field (Li and Mccallum, 2004 are most common. Combinations of different ML approaches are also used. Srihari et al. (2000) combines Maximum Entropy, Hidden Markov Model and handcrafted rules to build an NER system. NER systems use gazetteer lists for identifying names. Both the linguistic approach (Grishman, 1995; Wakao et al., 1996) and the ML based approach (Borthwick, 1999; Srihari et al., 2000) use gazetteer lists.",
"cite_spans": [
{
"start": 236,
"end": 261,
"text": "Model (Bikel et al., 1997",
"ref_id": null
},
{
"start": 262,
"end": 297,
"text": "), Maximum Entropy (Borthwick, 1999",
"ref_id": null
},
{
"start": 298,
"end": 348,
"text": "), Conditional Random Field (Li and Mccallum, 2004",
"ref_id": null
},
{
"start": 421,
"end": 442,
"text": "Srihari et al. (2000)",
"ref_id": "BIBREF9"
},
{
"start": 619,
"end": 635,
"text": "(Grishman, 1995;",
"ref_id": "BIBREF4"
},
{
"start": 636,
"end": 655,
"text": "Wakao et al., 1996)",
"ref_id": "BIBREF11"
},
{
"start": 682,
"end": 699,
"text": "(Borthwick, 1999;",
"ref_id": "BIBREF1"
},
{
"start": 700,
"end": 721,
"text": "Srihari et al., 2000)",
"ref_id": "BIBREF9"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Machine Learning based approaches.",
"sec_num": "2."
},
{
"text": "The linguistic approach uses hand-crafted rules which needs skilled linguistics. Some recent approaches try to learn context patterns through ML which reduce amount of manual labour. Talukder et al.(2006) combined grammatical and statistical techniques to create high precision patterns specific for NE extraction. An approach to lexical pattern learning for Indian languages is described by Ekbal and Bandopadhyay (2007) . They used seed data and annotated corpus to find the patterns for NER.",
"cite_spans": [
{
"start": 183,
"end": 204,
"text": "Talukder et al.(2006)",
"ref_id": null
},
{
"start": 392,
"end": 421,
"text": "Ekbal and Bandopadhyay (2007)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Machine Learning based approaches.",
"sec_num": "2."
},
{
"text": "The NER task for Hindi has been explored by Cucerzan and Yarowsky in their language independent NER work which used morphological and contextual evidences (Cucerzan and Yarowsky, 1999) . They ran their experiment with 5 languages -Romanian, English, Greek, Turkish and Hindi. Among these the accuracy for Hindi was the worst. For Hindi the system achieved 41.70% f-value with a very low recall of 27.84% and about 85% precision. A more successful Hindi NER system was developed by Wei Li and Andrew Mccallum (2004) using Conditional Random Fields (CRFs) with fea-ture induction. They were able to achieve 71.50% f-value using a training set of size 340k words. In Hindi the maximum accuracy is achieved by (Kumar and Bhattacharyya, 2006) . Their Maximum Entropy Markov Model (MEMM) based model gives 79.7% f-value.",
"cite_spans": [
{
"start": 155,
"end": 184,
"text": "(Cucerzan and Yarowsky, 1999)",
"ref_id": "BIBREF2"
},
{
"start": 481,
"end": 514,
"text": "Wei Li and Andrew Mccallum (2004)",
"ref_id": "BIBREF6"
},
{
"start": 706,
"end": 737,
"text": "(Kumar and Bhattacharyya, 2006)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Machine Learning based approaches.",
"sec_num": "2."
},
{
"text": "We have used a Maximum Entropy model to build the NER in Hindi. MaxEnt is a flexible statistical model which assigns an outcome for each token based on its history and features. MaxEnt computes the probability p(o|h) for any o from the space of all possible outcomes O, and for every h from the space of all possible histories H. A history is all the conditioning data that enables one to assign probabilities to the space of outcomes. In NER, history can be viewed as all information derivable from the training corpus relative to the current token. The computation of p(o|h) in MaxEnt depends on a set of features, which are helpful in making predictions about the outcome. The features may be binary-valued or multi-valued. For instance, one of our features is: the current token is a part of the surname list; how likely is it to be part of a person name. Formally, we can represent this feature as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Maximum Entropy Based Model",
"sec_num": "3"
},
{
"text": "f (h, o) =",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Maximum Entropy Based Model",
"sec_num": "3"
},
{
"text": "1 if w i in surname list and o = person 0 otherwise (1) Given a set of features and a training corpus, the MaxEnt estimation process produces a model in which every feature f i has a weight \u03b1 i . We can compute the conditional probability as (Pietra et al., 1997) :",
"cite_spans": [
{
"start": 242,
"end": 263,
"text": "(Pietra et al., 1997)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Maximum Entropy Based Model",
"sec_num": "3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "p(o|h) = 1 Z(h) i \u03b1 i f i (h,o) (2) Z(h) = o i \u03b1 i f i (h,o)",
"eq_num": "(3)"
}
],
"section": "Maximum Entropy Based Model",
"sec_num": "3"
},
{
"text": "So the conditional probability of the outcome is the product of the weights of all active features, normalized over the products of all the features. For our development we have used a Java based opennlp MaxEnt toolkit 1 to get the probability values of a word belonging to each class. That is, given a sequence of words, the probability of each class is obtained for each word. To find the most probable tag corresponding to each word of a sequence, we can choose the tag having the highest class conditional probability value. But this method is not good as it might result in an inadmissible output tag.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Maximum Entropy Based Model",
"sec_num": "3"
},
{
"text": "Some tag sequences should never happen. To eliminate these inadmissible sequences we have made some restrictions. Then we used a beam search algorithm with a beam of length 3 with these restrictions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Maximum Entropy Based Model",
"sec_num": "3"
},
{
"text": "The training data for this task is composed of about 243K words which is collected from the popular daily Hindi newspaper \"Dainik Jagaran\". This corpus has been manually annotated and has about 16,482 NEs. In this development we have considered 4 types of NEs, these are P erson(P), Location(L), Organization(O) and Date(D). To recognize entity boundaries each name class N is subdivided into 4 sub-classes, i.e., N Begin, N Continue, N End, and N U nique. Hence, there are a total of 17 classes including 1 class for not-name. The corpus contains 6, 298 Person, 4, 696 Location, 3, 652 Organization and 1, 845 Date entities.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Maximum Entropy Based Model",
"sec_num": "3"
},
{
"text": "Machine learning approaches like MaxEnt, CRF etc. make use of different features for identifying the NEs. Orthographic features (like capitalization, decimal, digits), affixes, left and right context (like previous and next words), NE specific trigger words, gazetteer features, POS and morphological features etc. are generally used for NER. In English and some other languages, capitalization features play an important role as NEs are generally capitalized for these languages. Unfortunately this feature is not applicable for Hindi. Also Indian person names are more diverse, lots of common words having other meanings are also used as person names. These make difficult to develop a NER system on Hindi. Li and Mccallum (2004) used the entire word text, character n-grams (n = 2, 3, 4), word prefix and suffix of lengths 2, 3 and 4, and 24 Hindi gazetteer lists as atomic features in their Hindi NER. Kumar and Bhattacharyya (2006) used word features (suffixes, digits, special characters), context features, dictionary features, NE list features etc. in their MEMM based Hindi NER system. In the following we have discussed about the features we have identified and used to develop the Hindi NER system.",
"cite_spans": [
{
"start": 709,
"end": 731,
"text": "Li and Mccallum (2004)",
"ref_id": "BIBREF6"
},
{
"start": 906,
"end": 936,
"text": "Kumar and Bhattacharyya (2006)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Features for Hindi NER",
"sec_num": "4"
},
{
"text": "The features which we have identified for Hindi Named Entity Recognition are:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Description",
"sec_num": "4.1"
},
{
"text": "Static Word Feature: The previous and next words of a particular word are used as features. The previous m words (w i\u2212m ...w i\u22121 ) to next n words (w i+1 ...w i+n ) can be treated. During our experiment different combinations of previous 4 to next 4 words are used.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Description",
"sec_num": "4.1"
},
{
"text": "Context Lists: Context words are defined as the frequent words present in a word window for a particular class. We compiled a list of the most frequent words that occur within a window of w i\u22123 ...w i+3 of every NE class. For example, location context list contains the words like 'jAkara 2 ' (going to), 'desha' (country), 'rAjadhAnI' (capital) etc. and person context list contains 'kahA' (say), 'prdhAnama.ntrI' (prime minister) etc. For a given word, the value of this feature corresponding to a given NE type is set to 1 if the window w i\u22123 ...w i+3 around the w i contains at last one word from this list.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Description",
"sec_num": "4.1"
},
{
"text": "Dynamic NE tag: Named Entity tags of the previous words (t i\u2212m ...t i\u22121 ) are used as features.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Description",
"sec_num": "4.1"
},
{
"text": "First Word: If the token is the first word of a sentence, then this feature is set to 1. Otherwise, it is set to 0.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Description",
"sec_num": "4.1"
},
{
"text": "Contains Digit: If a token 'w' contains digit(s) then the feature ContainsDigit is set to 1. This feature is helpful for identifying company product names (e.g. 06WD1992), house number (e.g. C226) etc.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Description",
"sec_num": "4.1"
},
{
"text": "Numerical Word: For a token 'w' if the word is a numerical word i.e. a word denoting a number (e.g. eka (one), do (two), tina (three) etc.) then the feature N umW ord is set to 1.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Description",
"sec_num": "4.1"
},
{
"text": "Word Suffix: Word suffix information is helpful to identify the named NEs. Two types of suffix features have been used. Firstly a fixed length word suffix of the current and surrounding words are used as features. Secondly we compiled lists of common suffixes of person and place names in Hindi. For example, 'pura', 'bAda', 'nagara' etc. are location suffixes. We used two binary features corresponding to the lists -whether a given word has a suffix from the list.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Description",
"sec_num": "4.1"
},
{
"text": "Word Prefix: Prefix information of a word may be also helpful in identifying whether it is a NE. A fixed length word prefix of current and surrounding words are treated as a features.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Description",
"sec_num": "4.1"
},
{
"text": "Parts-of-Speech (POS) Information: The POS of the current word and the surrounding words may be useful feature for NER. We have access to a Hindi POS pagger developed at IIT Kharagpur which has an accuracy about 90%. The tagset of the tagger contains 28 tags. We have used the POS values of the current and surrounding tokens as features.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Description",
"sec_num": "4.1"
},
{
"text": "We realized that the detailed POS tagging is not very relevant. Since NEs are noun phrases, the noun tag is very relevant. Further the postposition following a name may give a clue to the NE type. So we decided to use a coarse-grained tagset with only three tags -nominal (Nom), postposition (PSP) and other (O).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Description",
"sec_num": "4.1"
},
{
"text": "The POS information is also used by defining several binary features. An example is the N omP SP binary feature. The value of this feature is defined to be 1 if the current token is nominal and the next token is a PSP.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Feature Description",
"sec_num": "4.1"
},
{
"text": "Lists of names of various types are helpful in name identification. We have compiled some specialized name lists from different web sources. But the names in these lists are in English, not in Hindi. So we have transliterated these English name lists to make them useful for our Hindi NER task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Enhancement using Gazetteer Feature",
"sec_num": "5"
},
{
"text": "For the transliteration we have build a 2-phase transliteration module. We have defined an intermediate alphabet containing 34 characters. English names are transliterated to this intermediate form using a map-table. Hindi strings are also transliterated to the intermediate alphabet form using a different map-table. For a English-Hindi string pair, if transliterations of the both strings are same, then we conclude that one string is the transliteration of the other. This transliteration module works with 91.59% accuracy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Enhancement using Gazetteer Feature",
"sec_num": "5"
},
{
"text": "Using the transliteration approach we have constructed 8 lists. Which are, month name and days of the week (40) 3 , organization end words list (92), person prefix words list (123), list of common locations (80), location names list (17,600), first names list (9722), middle names list (35), surnames list (1800).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Enhancement using Gazetteer Feature",
"sec_num": "5"
},
{
"text": "The lists can be used in name identification in various ways. One way is to check whether a token is in any list. But this approach is not good as it has some limitations. Some words may present in two or more gazetteer lists. For example, 'bangAlora' is in surnames list and also in location names list. Confusions arise to make decisions for these words. Some words are in gazetteer lists but sometimes these are used in text as not-name entity. For example, 'gayA' is in location list but sometimes the word is used as verb in text and makes confusion. These limitations might be reduced if the contexts are considered.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Enhancement using Gazetteer Feature",
"sec_num": "5"
},
{
"text": "We have used these gazetteer lists as features of MaxEnt. We have prepared several binary features which are defined as whether a given word is in a particular list. For example, a binary feature F irstN ame is 1 for a particular token 't' if 't' is in the first name list.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Enhancement using Gazetteer Feature",
"sec_num": "5"
},
{
"text": "Context patterns are helpful for identifying NEs. As manual identification of context patterns takes much manual labour and linguistic knowledge, we have developed a module for semi-automatically learning of context pattern. The summary of the context pattern learning module is given follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context Pattern based Features",
"sec_num": "6"
},
{
"text": "1. Collect some seed entities (E) for each class.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context Pattern based Features",
"sec_num": "6"
},
{
"text": "2. For each seed entity e in E, from the corpus find context string(C) comprised of n tokens before e, a placeholder for the class instance and n tokens after e. [We have used n = 3] This set of tokens form initial pattern.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context Pattern based Features",
"sec_num": "6"
},
{
"text": "3. Search the pattern in the corpus and find the coverage and precision.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context Pattern based Features",
"sec_num": "6"
},
{
"text": "4. Discard the patterns having low precision.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context Pattern based Features",
"sec_num": "6"
},
{
"text": "5. Generalize the patterns by dropping one or more tokens to increase coverage.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context Pattern based Features",
"sec_num": "6"
},
{
"text": "6. Find best patterns having good precision and coverage.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context Pattern based Features",
"sec_num": "6"
},
{
"text": "The quality of a pattern is measured by precision and coverage. Precision is the ratio of correct identification and the total identification, when the particular pattern is used to identify of NEs of a specific type from a raw text. Coverage is the amount of total identification. We have given more importance to precision and we have marked a pattern as ef f ective if the precision is more than 95%. The method is applied on an un-annotated text having 4887011 words collected from \"Dainik Jagaran\" and context patterns are learned. These context patterns are used as features of MaxEnt in the Hindi NER system. Some example patterns are: ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Context Pattern based Features",
"sec_num": "6"
},
{
"text": "We have evaluated the system using a blind test corpus of 25K words, which is distinct from the training corpus. The accuracies are measured in terms of the f-measure, which is the weighted harmonic mean of precision and recall. Here we can mention that we have evaluated the performance of the system on actual NEs. That means the system annotates the test data using 17 tags, similar to the training data. During evaluation we have merged the sub-tags of a particular entity to get a complete NEs and calculated the accuracies. At the end of section 7.1 we have also mentioned the accuracies if evaluated on the tags. A number of experiments are conducted considering various combinations of features to identify the best feature set for the Hindi NER task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Evaluation",
"sec_num": "7"
},
{
"text": "The baseline performance of the system without using gazetteer and context patterns are presented in Table 1 . They are summarized below.",
"cite_spans": [],
"ref_spans": [
{
"start": 101,
"end": 108,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Baseline",
"sec_num": "7.1"
},
{
"text": "While experimenting with static word features, we have observed that a window of previous two Similarly we have experimented with suffixes of different lengths and observed that the suffixes of length \u2264 2 gives the best result for the Hindi NER task. In using POS information, we have observed that the coarse-grained POS tagger information is more effective than the finer-grained POS values. A feature set, combining finer-grained POS values, surrounding words and previous NE tag, gives a f-value of 70.39%. But when the coarse-grained POS values are used instead of the finer-grained POS values, the f-value is increased to 74.16%. The most interesting fact we have observed that more complex features do not guarantee to achieve better results. For example, a feature set combined with current and surrounding words, previous NE tag and fixed length suffix information, gives a f-value 73.42%. But when prefix information are added the f-value decreased to 72.5%. The highest accuracy achieved by the system is 75.6% f-value without using gazetteer information and context patterns.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Baseline",
"sec_num": "7.1"
},
{
"text": "The results in Table 1 are obtained by evaluating on the actual NEs. But when the system is evaluated on the tags the f-value increases. For f6, the accuracy achieved on actual NEs is 75.6%, but if evaluated on tags, the value increased to 77.36%. Similarly, for f2, the accuracy increased to 75.91% if evaluated on tags. The reason is the NEs containing 3 or more words, are subdivided to N-begin, Ncontinue (1 or more) and N-end. So if there is an error in any of the subtags, the total NE becomes an error. We observed many cases where NEs are partially identified by the system, but these are considered as error during evaluation.",
"cite_spans": [],
"ref_spans": [
{
"start": 15,
"end": 22,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Baseline",
"sec_num": "7.1"
},
{
"text": "Next we add gazetteer and context patterns as features in our MaxEnt based NER system. In Table 2 we have compared the results after addition of gazetteer information and context patterns with previous results. While experimenting we have observed that gazetteer lists and context patterns are capable of increasing the performance of our baseline system. That is tested on all the baseline feature sets. In Table 2 the comparison is shown for only two features -f2 and f6 which are defined in Table 1 . It may be observed that the relative advantage of using both gazetteer and context patterns together over using them individually is not much. For example, when gazetteer information are added with f2, the fvalue is increased by 6.38%, when context patterns are added the f-value is increased by 6.64%., but when both are added the increment is 7.27%. This may be due to the fact that both gazetteer and context patterns lead to the same identifications. Using the comprehensive feature set (using gazetteer information and context patterns) the MaxEnt based NER system achieves the maximum f-value of 81.52%. ",
"cite_spans": [],
"ref_spans": [
{
"start": 90,
"end": 97,
"text": "Table 2",
"ref_id": "TABREF3"
},
{
"start": 408,
"end": 415,
"text": "Table 2",
"ref_id": "TABREF3"
},
{
"start": 494,
"end": 501,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "Using Gazetteer Lists and Context Patterns",
"sec_num": "7.2"
},
{
"text": "We have shown that our MaxEnt based NER system is able to achieve a f-value of 81.52%, using a hybrid set of features including traditional NER features augmented with gazetteer lists and extracted context patterns. The system outperforms the existing NER systems in Hindi.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "8"
},
{
"text": "Feature selection and feature clustering might lead to further improvement of performance and is under investigation.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "8"
},
{
"text": "www.maxent.sourceforge.net.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "All Hindi words are written in italics using the 'Itrans' transliteration.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "The italics integers in brackets indicate the size of the lists.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "The work is partially funded by Microsoft Research India.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgement",
"sec_num": "9"
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Nymble: A High Performance Learning Name-finder",
"authors": [
{
"first": "Bikel",
"middle": [],
"last": "Daniel",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "",
"suffix": ""
},
{
"first": "Miller",
"middle": [],
"last": "Scott",
"suffix": ""
},
{
"first": "Schwartz",
"middle": [],
"last": "Richard",
"suffix": ""
},
{
"first": "Weischedel",
"middle": [],
"last": "Ralph",
"suffix": ""
}
],
"year": 1997,
"venue": "Proceedings of the Fifth Conference on Applied Natural Language Processing",
"volume": "",
"issue": "",
"pages": "194--201",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Bikel Daniel M., Miller Scott, Schwartz Richard and Weischedel Ralph. 1997. Nymble: A High Perfor- mance Learning Name-finder. In Proceedings of the Fifth Conference on Applied Natural Language Pro- cessing, pages 194-201.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "A Maximum Entropy Approach to Named Entity Recognition",
"authors": [
{
"first": "Borthwick",
"middle": [],
"last": "Andrew",
"suffix": ""
}
],
"year": 1999,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Borthwick Andrew. 1999. A Maximum Entropy Ap- proach to Named Entity Recognition. Ph.D. thesis, Computer Science Department, New York University.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Language Independent Named Entity Recognition Combining Morphological and Contextual Evidence",
"authors": [
{
"first": "Cucerzan",
"middle": [],
"last": "Silviu",
"suffix": ""
},
{
"first": "Yarowsky",
"middle": [],
"last": "David",
"suffix": ""
}
],
"year": 1999,
"venue": "Proceedings of the Joint SIGDAT Conference on EMNLP and VLC 1999",
"volume": "",
"issue": "",
"pages": "90--99",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Cucerzan Silviu and Yarowsky David. 1999. Language Independent Named Entity Recognition Combining Morphological and Contextual Evidence. In Proceed- ings of the Joint SIGDAT Conference on EMNLP and VLC 1999, pages 90-99.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Lexical Pattern Learning from Corpus Data for Named Entity Recognition",
"authors": [
{
"first": "A",
"middle": [],
"last": "Ekbal",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Bandyopadhyay",
"suffix": ""
}
],
"year": 2007,
"venue": "Proceedings of International Conference on Natural Language Processing (ICON)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ekbal A. and Bandyopadhyay S. 2007. Lexical Pattern Learning from Corpus Data for Named Entity Recog- nition. In Proceedings of International Conference on Natural Language Processing (ICON), 2007.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "The New York University System MUC-6 or Where's the syntax?",
"authors": [
{
"first": "Grishman",
"middle": [],
"last": "Ralph",
"suffix": ""
}
],
"year": 1995,
"venue": "Proceedings of the Sixth Message Understanding Conference",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Grishman Ralph. 1995. The New York University Sys- tem MUC-6 or Where's the syntax? In Proceedings of the Sixth Message Understanding Conference.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "Named Entity Recognition in Hindi using MEMM",
"authors": [
{
"first": "N",
"middle": [],
"last": "Kumar",
"suffix": ""
},
{
"first": "Bhattacharyya",
"middle": [],
"last": "Pushpak",
"suffix": ""
}
],
"year": 2006,
"venue": "Technical Report, IIT",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kumar N. and Bhattacharyya Pushpak. 2006. Named Entity Recognition in Hindi using MEMM. In Techni- cal Report, IIT Bombay, India..",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Rapid Development of Hindi Named Entity Recognition using Conditional Random Fields and Feature Induction (Short Paper)",
"authors": [
{
"first": "Li",
"middle": [],
"last": "Wei",
"suffix": ""
},
{
"first": "Mccallum",
"middle": [],
"last": "Andrew",
"suffix": ""
}
],
"year": 2004,
"venue": "ACM Transactions on Computational Logic",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Li Wei and McCallum Andrew. 2004. Rapid Develop- ment of Hindi Named Entity Recognition using Con- ditional Random Fields and Feature Induction (Short Paper). ACM Transactions on Computational Logic.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Internal and external evidence in the identification and semantic categorization of proper names",
"authors": [
{
"first": "D",
"middle": [],
"last": "Mcdonald",
"suffix": ""
}
],
"year": 1996,
"venue": "Corpus Processing for Lexical Acquisition",
"volume": "",
"issue": "",
"pages": "21--39",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "McDonald D. 1996. Internal and external evidence in the identification and semantic categorization of proper names. In B. Boguraev and J. Pustejovsky, editors, Corpus Processing for Lexical Acquisition, pages 21- 39.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Inducing features of random fields",
"authors": [
{
"first": "Pietra",
"middle": [
"Stephen"
],
"last": "Della",
"suffix": ""
},
{
"first": "Pietra",
"middle": [
"Vincent"
],
"last": "Della",
"suffix": ""
},
{
"first": "Lafferty",
"middle": [],
"last": "John",
"suffix": ""
}
],
"year": 1997,
"venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
"volume": "19",
"issue": "4",
"pages": "380--393",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Pietra Stephen Della, Pietra Vincent Della and Lafferty John. 1997. Inducing features of random fields. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 19(4):380-393.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "A Hybrid Approach for Named Entity and Sub-Type Tagging",
"authors": [
{
"first": "R",
"middle": [],
"last": "Srihari",
"suffix": ""
},
{
"first": "C",
"middle": [],
"last": "Niu",
"suffix": ""
},
{
"first": "W",
"middle": [],
"last": "Li",
"suffix": ""
}
],
"year": 2000,
"venue": "Proceedings of the sixth conference on Applied natural language processing",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Srihari R., Niu C. and Li W. 2000. A Hybrid Approach for Named Entity and Sub-Type Tagging. In Proceed- ings of the sixth conference on Applied natural lan- guage processing.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "A context pattern induction method for named entity extraction",
"authors": [
{
"first": "Talukdar",
"middle": [],
"last": "Pratim",
"suffix": ""
},
{
"first": "P",
"middle": [],
"last": "Brants",
"suffix": ""
},
{
"first": "T",
"middle": [],
"last": "Liberman",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Pereira",
"suffix": ""
},
{
"first": "F",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 2006,
"venue": "Proceedings of the Tenth Conference on Computational Natural Language Learning (CoNLL-X)",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Talukdar Pratim P., Brants T., Liberman M., and Pereira F. 2006. A context pattern induction method for named entity extraction. In Proceedings of the Tenth Conference on Computational Natural Lan- guage Learning (CoNLL-X).",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Evaluation of an algorithm for the recognition and classification of proper names",
"authors": [
{
"first": "T",
"middle": [],
"last": "Wakao",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Gaizauskas",
"suffix": ""
},
{
"first": "Y",
"middle": [],
"last": "Wilks",
"suffix": ""
}
],
"year": 1996,
"venue": "Proceedings of COLING-96",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Wakao T., Gaizauskas R. and Wilks Y. 1996. Evaluation of an algorithm for the recognition and classification of proper names. In Proceedings of COLING-96.",
"links": null
}
},
"ref_entries": {
"TABREF1": {
"num": null,
"type_str": "table",
"content": "<table><tr><td>Feature</td><td>Class</td><td>F-value</td></tr><tr><td/><td>PER</td><td>63.33</td></tr><tr><td/><td>LOC</td><td>69.56</td></tr><tr><td>f1 = Word, NE Tag</td><td>ORG</td><td>58.58</td></tr><tr><td/><td>DAT</td><td>91.76</td></tr><tr><td/><td colspan=\"2\">TOTAL 69.64</td></tr><tr><td/><td>PER</td><td>69.75</td></tr><tr><td/><td>LOC</td><td>75.8</td></tr><tr><td>f2 = Word, NE Tag,</td><td>ORG</td><td>59.31</td></tr><tr><td>Suffix (\u2264 2)</td><td>DAT</td><td>89.09</td></tr><tr><td/><td colspan=\"2\">TOTAL 73.42</td></tr><tr><td/><td>PER</td><td>70.61</td></tr><tr><td/><td>LOC</td><td>71</td></tr><tr><td>f3 = Word, NE Tag,</td><td>ORG</td><td>59.31</td></tr><tr><td>Suffix (\u2264 2), Prefix</td><td>DAT</td><td>89.09</td></tr><tr><td/><td colspan=\"2\">TOTAL 72.5</td></tr><tr><td/><td>PER</td><td>70.61</td></tr><tr><td/><td>LOC</td><td>75.8</td></tr><tr><td>f4 = Word, NE Tag,</td><td>ORG</td><td>60.54</td></tr><tr><td>Digit, Suffix (\u2264 2)</td><td>DAT</td><td>93.8</td></tr><tr><td/><td colspan=\"2\">TOTAL 74.26</td></tr><tr><td/><td>PER</td><td>64.25</td></tr><tr><td/><td>LOC</td><td>71</td></tr><tr><td>f5 = Word, NE Tag, POS</td><td>ORG</td><td>60.54</td></tr><tr><td/><td>DAT</td><td>89.09</td></tr><tr><td/><td colspan=\"2\">TOTAL 70.39</td></tr><tr><td/><td>PER</td><td>72.26</td></tr><tr><td>f6 = Word, NE Tag,</td><td>LOC</td><td>78.6</td></tr><tr><td>Suffix (\u2264 2), Digit,</td><td>ORG</td><td>51.36</td></tr><tr><td>N omP SP</td><td>DAT</td><td>92.82</td></tr><tr><td/><td colspan=\"2\">TOTAL 75.6</td></tr></table>",
"html": null,
"text": "Table 1: F-values for different features words to next two words (W i\u22122 ...W i+2 ) gives best results. But when several other features are combined then single word window (W i\u22121 ...W i+1 ) performs better."
},
"TABREF3": {
"num": null,
"type_str": "table",
"content": "<table/>",
"html": null,
"text": "F-values for different features with gazetteers and context patterns"
}
}
}
} |